Question

I have a multiclass svm classification(6 class). I would like to classify it using LIBSVM. The following are the ones that i have tried and i have some questions regarding them.

Method1( one vs one):

model = svmtrain(TrainLabel, TrainVec, '-c 1 -g 0.00154 -b 0.9');
[predict_label, accuracy, dec_values] = svmpredict(TestLabel, TestVec, model);

Two questions about this method: 1) is that all i need to do for multiclass problem
2) what value should it be for n in ‘-b n’. I m not sure

Method 2( one vs rest):

u=unique(TrainLabel); 
N=length(u); 
if(N>2)    
    itr=1;    
    classes=0;   
    while((classes~=1)&&(itr<=length(u)))   
        c1=(TrainLabel==u(itr));    
        newClass=double(c1); 
        tst = double((TestLabel == itr));
        model = svmtrain(newClass, TrainVec, '-c 1 -g 0.00154');  
        [predict_label, accuracy, dec_values] = svmpredict(tst, TestVec, model);    
        itr=itr+1;   
    end
    itr=itr-1;
end

For the second method,how do I attach classification scores. I am not able to do voting.

Besides that,these are the two methods I have tried. Which method is better?

Would like to hear some comments. Please correct me if I am wrong.

Answer 1

Regarding the ‘-b’ parameter, in the LIBSVM README it says:

-b probability_estimates: whether to train a SVC or SVR model for probability estimates, 0 or 1 (default 0)

Therefore, you should specify ‘-b 1’ if you want the trained model to return class probabilities, and ‘-b 0’ if you don’t. You only need to call svmtrain once. Also, if you specify ‘-b 1’ for training, you must also specify it for prediction.